Charles Explorer logo
🇬🇧

Neural networks with Wavelet based denoising layer : application to Central European stock market forecasting

Publication at Faculty of Social Sciences |
2008

Abstract

Traditional prediction methods for time series often restrict on linear regression analysis, exponential smoothing, and ARMA. These methods generally produce reasonable prediction results for stationary random time series of linear systems.

In the recent decades, development in econometrics resulted also in methods which are capable of forecasting more complex systems, such as Wavelet decomposition or Neural Networks. These methods proved to better explain the complex stock market behavior.

In this paper we apply neural network with wavelet denoising layer method for forecasting of Central European Stock Exchanges, namely Prague, Budapest and Warsaw. Hard threshold denoising with Daubechies 6 wavelet filter and three level decomposition is used to denoise the stock index returns, and two-layer feed-forward neural network with Levenberg-Marquardt learning algorithm is used for modeling.

The results show that wavelet network structure is able to approximate the underlying process of considered stock markets better that multilayered neural network architecture without using wavelets. Further on we discuss the impact of structural changes of the market on forecasting accuracy on the daily stock market data.

Stock markets change their structure rapidly with changing agent sentiment structure. These changes then have great impact on the prediction accuracy.